Thanks for the remarks, I don’t understand well enough yet how the scan with rng works in the pymc source. I was trying to do without the potential because the Muse package doesn’t support it, and I wanted to compare the inference algorithms for centered and non-centered cases. I’ll keep at it and come back later.
edit For instance this is clearly the centered parameterization. But the scan implementation above in the rv_op method seems redundant and potentially non-centered. The posterior geometries between centered and non-centered are different, hence my curiousity.